Skip to content

Latest commit

 

History

History
48 lines (37 loc) · 1.68 KB

README.md

File metadata and controls

48 lines (37 loc) · 1.68 KB
Raw

image-quality-assessment

Private 1st code of AI competition "2023 Samsung AI Challenge : Image Quality Assessment" hosted by Samsung AIT & Dacon

Overview

There are two tasks for this competition, image quality assesment and image captioning.

image quality assessment

  1. Extract features from image with pretrained neural network
  2. training autogluon model with extracted features. I extracted features from 4 pretraiend models with 2 version of input. (total 8)
  • Models: ViT Huge, Swin Transformer 1K, Swin Transformer 22K, NFNet F5.
  • Inputs: original image, horizontal flipped image.

Final result is average ensemble of them.

image captioning

I utilized ExpansionNet V2, which outperforms among other arhcitectures.
Check src/caption_src/ README.md for details.

Get Started

Feature extractin for image quality assessment

python ./src/image_src/feature_extraction.py --csv_dir <csv_dir> --root_dir <root_dir> --save_dir <save_dir> --model_name <model_name> --img_size <img_size>
  • ViT Huge - model_name: ViT_H, img_size: 518
  • Swin 1k - model_name: Swin, img_size: 384
  • Swin 22k - model_name: Swin, img_size: 192
  • NFNet F5 - model_name: NFNetF5, img_size: 544

flip argument is optional, and every model requries pretrained weight. (state_dict_dir, nf_weight_dir)

For extracted feature files, please don't hesitate to reach out to me via Gmail.

Run autogluon model

For training,

python ./src/image_src/autogluon_reressor.py --data <data_dir>

For inference,

python ./src/image_src/autogluon_inference.py

Sources

I utilized code for image captioning repository below (ExpansionNet V2). https://github.com/jchenghu/ExpansionNet_v2